the short answer
Score is the Jev primitive for rating state against caller-defined ordered levels. It returns a probability for each level, confidence and a probability-weighted expected score that can fall between level indices. Each level should describe a concrete semantic situation. Jev 1.13 Score is not intended for arithmetic, counting or reconstructing an exact numeric value.
- Question shape
- Ordered descriptive criteria
- Response
- Expected score, probabilities, legend and confidence
- Best use
- Severity, quality or fit on a semantic rubric
- Not for
- Exact numbers or calculation
Write Levels That Stand on Their Own
A useful rubric describes observable situations at each ordered level. “Low, medium, high” leaves the boundary implicit. “No user impact”, “recoverable delay” and “irreversible user impact” gives the model and reviewers a shared interpretation. Positions begin at zero and determine the numeric score.
The returned score is an expectation over the level probabilities, not a directly observed quantity. A 1.6 does not mean the underlying event has a measurable magnitude of 1.6. Preserve the full distribution and legend beside it.
Use Score for Ordered Decisions, Then Apply Policy in Code
An application can combine a semantic Score with deterministic vetoes or weights. For example, any confirmed permission violation may require review regardless of a lower aggregate quality score. Do not average away a serious independent condition; ask it separately.
Example: Define Urgency with Observable Levels
These levels describe one progression: increasing operational urgency. Do not mix unrelated dimensions into the same scale, such as customer tier, sentiment and security risk. Ask separate questions and combine them explicitly when the workflow needs all three. The Score-versus-Choice guide gives the simpler selection rule.
| Level | Meaning |
|---|---|
| Routine | No stated deadline, blocked workflow or material consequence |
| Time-sensitive | A deadline or degraded workflow exists, but a safe workaround is available |
| Critical | A critical process is blocked or a near-term material deadline is at risk |
The Same Average Can Hide Different Uncertainty
Imagine a three-level Score whose positions are 0, 1 and 2. A distribution concentrated on the middle level and a distribution split evenly between the lowest and highest levels can both average to 1.0. The first suggests a clear middle judgment; the second suggests conflicting evidence. Storing only the average erases that distinction.
Use the level probabilities and confidence when deciding whether to automate, review or gather more evidence. Validate those bands on labeled outcomes rather than assigning a universal cutoff. See confidence versus probability and threshold selection.
Reproduce the Expected Score from the Distribution
The expected score is 1.25, but no rubric level named 1.25 exists. It summarizes probability mass across the three positions; it does not say the incident has 1.25 units of urgency. Preserve the legend and probabilities so downstream code and reviewers can see that critical still carries 0.35 probability.
Level positions encode order and spacing chosen by the caller. Moving from three to five levels or inserting a new level changes the numeric meaning even if the words look similar. Do not compare expected scores across rubric versions without an explicit migration study.
| Level position | Rubric label | Probability | Contribution |
|---|---|---|---|
| 0 | Routine | 0.10 | 0 × 0.10 = 0.00 |
| 1 | Time-sensitive | 0.55 | 1 × 0.55 = 0.55 |
| 2 | Critical | 0.35 | 2 × 0.35 = 0.70 |
| Total | Expected score | 1.00 | 1.25 |
Test Adjacent Boundaries and Distant Mistakes
Overall agreement hides distance. Confusing adjacent levels may be tolerable; jumping from routine to critical may overload an escalation queue, while the reverse can miss a serious incident. Define a cost matrix before selecting thresholds and report the full confusion matrix. The dataset guide explains how to preserve difficult boundary cases.
| Test case | What it reveals |
|---|---|
| Routine versus time-sensitive near the stated deadline | Whether adjacent criteria have a usable boundary |
| Time-sensitive versus critical with and without a workaround | Whether decisive evidence moves probability in the intended direction |
| Same incident with customer tier removed | Whether an irrelevant field is steering urgency |
| Paraphrased but equivalent report | Whether wording changes produce unstable levels |
| Routine labeled as critical | A distant, potentially expensive ordinal error |
Good Score Uses
- Rate support-ticket urgency against operational response levels.
- Evaluate handoff quality against evidence-based descriptions.
- Rank candidates on one semantic dimension while retaining uncertainty.
- Measure agent-session severity before grouping failures.
Evaluate Ordinal Errors, Not Only the Average Score
A one-level miss and a two-level miss usually have different consequences. Define those costs before selecting action bands. The evaluator scorecard covers held-out labels, uncertainty and operational utility.
- Report the confusion matrix across rubric levels.
- Measure how far wrong predictions move on the ordered scale.
- Inspect the complete distribution for split or diffuse cases.
- Evaluate each operational slice and high-consequence boundary.
- Repeat after any rubric, state, model or provider change.
Treat Every Rubric Change as a New Instrument
Changing a label, definition, order or number of levels changes what the Score measures. Freeze a labeled regression set, run both rubric versions, compare level confusion and probability calibration, then inspect cases that cross an operational boundary. Do not tune on the final test set or map old and new decimals by assumption.
Store rubric version, state-projection version, requested and resolved model, provider, complete distribution, expected score and derived action. That record supports replay when an alias moves or a threshold changes. The regression-testing guide covers tolerances and promotion gates.
FAQ
Why does Jev Score return decimals?
The score is the probability-weighted average of level positions. Probability spread across levels can therefore produce a fractional value.
Can Score extract an exact number?
No. TypeSafe warns that Jev 1.13 is weak at numeric calibration and should not reconstruct an exact magnitude between rubric levels.
How many Score levels should I use?
Use the smallest set that describes operationally distinct outcomes. Each level needs a standalone definition and representative validation examples.
Should I keep Score probabilities?
Yes. The expected score alone hides whether the distribution is concentrated or split across distant levels.
Sources
Checked against the sources below on September 22, 2026. Model versions, prices and limits change.
- TypeSafe AI docs: Primitives
- TypeSafe AI docs: Confidence
- TypeSafe AI docs: Jev 1.13 jaggedness
- scikit-learn: Classification metrics and confusion matrices